CN107564097A - A kind of remains of the deceased three-dimensional rebuilding method based on direct picture - Google Patents
A kind of remains of the deceased three-dimensional rebuilding method based on direct picture Download PDFInfo
- Publication number
- CN107564097A CN107564097A CN201710793207.2A CN201710793207A CN107564097A CN 107564097 A CN107564097 A CN 107564097A CN 201710793207 A CN201710793207 A CN 201710793207A CN 107564097 A CN107564097 A CN 107564097A
- Authority
- CN
- China
- Prior art keywords
- face
- dimensional
- human face
- model
- characteristic point
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Processing Or Creating Images (AREA)
Abstract
A kind of remains of the deceased three-dimensional rebuilding method based on direct picture, it is characterised in that：A kind of single picture three-dimensional facial reconstruction technology based on statistical model.This method is in advance based on three-dimensional face database and establishes three-dimensional face statistical model, and obtains two-dimension human face characteristic point parameter model using SDM Algorithm for Training based on two-dimension human face database.Based on individual front shine carry out human face rebuilding when, first by two-dimension picture characteristic point parameter model extract human face characteristic point；Then, according to three-dimensional face statistical model, optimization is iterated to energy function using Studying factors self-adaption gradient descent method, statistical model parametrization vector is obtained, that is, obtains three-dimensional face model corresponding to two-dimension human face picture.As a result show, using method proposed by the present invention reconstruct come face there is higher similarity.
Description
Technical field
Face reconstruction techniques of the invention based on direct picture, employ the statistical model based on 3D human face scanning datas storehouse
The method being combined with parameter optimization, full-automatic generation can be directly used for the three-dimensional face model of 3D printing, be mainly used in something lost
Repair in honorable portion.
Background technology
Remains cosmetic surgery：Remains lift face, farewell are that people-oriented basic concepts and performance people are embodied during funeral and interment
The important step of class spiritual civilization development, is mainly completed by remains beautician.Due to the uniqueness of face, uniqueness, tradition is lost
Hold and repair to mould by hand, it is desirable to be painted, made up based on sculpture, material is mainly plasticine, gypsum, greasy filth etc.,
Cosmetic outcome is different because of the individual skill ability of anti-corrosion beautician.More serious or defect remains are deformed for Head And Face,
Can only photo be relied on to carry out refigure, just like sculpture is, it is necessary to high aesthetic and manual ability, even having abundant
The anti-corrosion beautician of experience, also can not necessarily reduce the deceased's original appearance, repair distortion, it is time-consuming it is extremely long, can not repair etc. and to be all
The problem of being difficult to overcome at present.In addition, anti-corrosion lift face is funeral worker contacted with remains most closely, health and epidemic prevention pressure it is maximum,
Technical requirements highest link.Therefore, the full-automatic generation technique of three-dimensional face model, it is greatly improved the work of remains beautician
Make condition and operating efficiency, reduce manual operation difficulty and safety and sanitation risk.
Face reconstruction techniques based on direct picture：Had based on the three-dimensional reconstruction of single picture in life extensively
Application prospect, also have very high technical difficulty.Three-dimensional facial reconstruction technology based on direct picture is broadly divided into based on machine
The method of study and the method based on statistical model.Blanz and Vetter research group is proposed based on 3-d deformable mould
The method for reconstructing of type, a 3-d deformable faceform, including three-dimensional shape model are obtained (with about by statistical learning first
70000 vertex representations) and corresponding 2 d texture model, then make deformable model alignment specific using Newton iteration method
Facial image, and then obtain the threedimensional model of the face, this method can obtain more accurate 3D shape and texture
Information, and it has been successfully applied to three-dimensional face identification.Tran and Hassner et al. are constructed using deep neural network
One system that three-dimensional face can be generated from single picture, method of this method than blanz more quickly and efficiently, can be with
It is used very advantageously in the system higher to requirement of real-time, such as video conference etc..
The content of the invention
The purpose of this method is to overcome a kind of remains of the deceased three-dimensional rebuilding method based on direct picture of prior art defect.
The object of the present invention is achieved like this：There are the remains of major injury for Head And Face, can be by a face just
Face photo automatically generates the three-dimensional face model that can be directly used for 3D printing.Feature is：Possess independent intellectual property right, improve
A kind of face statistical model generation method based on 3D human face scanning datas storehouse, propose a kind of face characteristic side of automatically extracting
Method；Later stage modification adjustment can be carried out to generation model；There are a variety of export forms such as mask, formpiston, former for different demands；And
System is easy to use, full-featured flexible.
Particular technique route is as follows：
(1) by establishing 2D the and 3D databases of face, the number of people, the parameter model and statistical model of face, the number of people are obtained；
(2) features of human face images is identified based on the parameter model of face；
(3) human face rebuilding based on image is completed based on face statistical model；
(4) reparation of the number of people is completed based on number of people statistical model, and utilizes the number of people symmetrical and number of people priori,
It is aided with the reparation operations such as mirror image, local modification.
Brief description of the drawings
Fig. 1 is the remains of the deceased three-dimensional rebuilding method overall flow frame diagram based on direct picture
Embodiment
1 master-plan
Face reconstruction techniques of this method based on direct picture, the statistical model based on 3D human face scanning datas storehouse of use
The method being combined with parameter optimization, Fig. 1 give the 3D human face rebuilding frameworks based on single image, and its global design is as follows：
(1) 3D human face scanning datas storehouse is based on, generates 3D face nets that one group of feature is alignd and that topological structure is consistent
Lattice model, using shape Statistics analysis method, the statistical model for establishing 3D faces (represents the number of people based on PCA probabilistic model
The statistics variations of portion's shape)；
(2) to a given face full face, lineup's face characteristic point is automatically extracted, with this group of characteristic point to statistics
The form parameter and projection matrix parameters of model optimize, and obtain the optimal 3D faceforms of given photo.
2 face databases
This project establishes the statistical model of 3D faces using BJUT-3D face databases.The large-scale Chinese faces of BJUT-3D
500 Chinese's three-dimensional faces are contained in database at present, men and women's face respectively accounts for half, and its data is public with CyberWare
3030RGB/PS laser scanners are taken charge of to obtain.Face keeps neutral expression during scanning, and does not wear glasses and jewelry.Scanning light source makes
It is scanner automatic light source, the light source can be with the photoenvironment of simulated environment light.
3 data predictions
Due to scanner when scanning face because a variety of causes can not obtain perfect three-dimensional face data, therefore doing
It is necessary completely into pretreatment is carried out to the data of acquisition before three-dimensional data base.Its step mainly includes data smoothing,
Remove burr, filling cavity (data of missing) and calibration coordinate (all unifying under human body head anatomical coordinates system) and three
Dimension cutting etc., finally obtain the 3D faceforms of the standardization of coordinate alignment.
The isomorphism mess generation of 4 features alignment
The statistical analysis of shape requires that all faces in all face 3D databases in 3D databases can use one
Shape vector and a texture represent.This needs to carry out feature mark (identification) to 3D faces, carries out feature alignment simultaneously
Generate the face triangle gridding of homogeneity.The isomorphism grid model of feature alignment refers to that one group of mesh topology is consistent, characteristic point
The triangle grid model of alignment.
After the homogeneity triangle gridding of each face generation feature alignment, each face can use a shape vector and one
Texture is as follows to represent difference：
Si=(x1, y1, z1, x2..., xn, yn, zn)T
Ti=(R1, G1, B1, R2..., Rn, Gn, Bn)T
Wherein 1≤i≤m, m are face number, and n is the points of each three-dimensional face.Shape vector SiIn element face on
Coordinate value x, y, the z of each point, and texture TiFor the RGB color value of each point.
5 face statistical models are established
After the isomorphism grid model for generating feature alignment is all established in all m training samples (face), it is possible to establish
Face statistical model.According to the shape vector and texture of each faceforms of m, PCA (principle is utilized
Component analysis, PCA) it can extract out shape eigenvectors and texture feature vector.
BJUT-3D face databases have used 200 faces (100 males and 100 women) to be used as sample (i.e. m=
200).Include average shape vector sum average texture vector using the PCA statistical models for analyzing to obtain, and corresponding to them
Characteristic vector.
PCA is separately to a series of shape vector S being made up of m sample faceiWith texture TiIn, i=
1 ..., m.For shape, a data matrix L=(l is defined1, l2..., lm)：
Wherein S is the average value of shape vector, and C is its covariance matrix：
According to a series of shape eigenvectors s of covariance matrix C1, s2..., it can similarly obtain a series of texture
Characteristic vector t1, t2..., characteristic vector forms one group of orthogonal basis, i.e.,：
Change shape and texture coefficients αiAnd βi, with regard to that can obtain different three-dimensional faces, these faces are likely to be former data
It is not present in storehouse.6 face characteristic parameter models are established and facial features localization
This technology realizes facial feature points detection using SDM algorithms, and the process is divided into two steps of training and detection：
(1) SDM model trainings
SDM needs to obtain R from training setkAnd bk.During training, face picture collection is { Ii, demarcated by hand per pictures
Characteristic point isInitial characteristicses point per pictures is X0, such facial modeling reformed into a linear regression and asked
Topic, the target of regression problem is from X0ArriveThe step-length of iteration, the input feature vector of this regression problem is exactly in X0The SIFT at place is special
Levy φ0.So, apply mechanically linear regression problem and can be obtained by object function：
R has thus been obtained from training set0And b0.Here it is to have obtained the coefficients R of first time iteration0And b0, identical
Method can obtain the coefficients R of kth time iterationkAnd bk
In this manner it is possible to the parameter R of each iteration is obtained from training setkAnd bk。
(2) SDM Face datections
P feature point coordinates in picture is represented, h is the Nonlinear feature extraction letter at each characteristic point
Number, feature used herein are SIFT features, that is, each characteristic point will extract the SIFT feature of 128 dimensions, so h (d (x))
∈128*1.It is by the optimal solution manually demarcated The SIFT feature of place's extraction can be designated asThus, people
Face characteristic point detection can be regarded as, ask and cause the minimum Δ x of lower array function：
Need exist for doing, obtain a series of direction and step delta x, by characteristic point from initial value X0Converge to optimal solutionHere by using the continuous iterative formula X of SDM algorithmsk=Xk-1+Rk-1φk-1+bk-1And optimal solution is finally given, in formula
RkAnd bkExactly previous step trains to obtain.
7 generation faces
After characteristic point demarcation above, face statistics mould can be adjusted to begin through the constraint of this group of characteristic point
Shape parameterization vector, and generate real three-dimensional face model.If three-dimensional statistical model obtained above is Smm, r characteristic point
Collect for { Pj, in order to be met the threedimensional model of face in picture as far as possible, it is only necessary to which the corresponding points allowed in statistical model are passed through
Parameter isCamera projective transformation after fall point on two dimensional surface and meet feature point set { P as far as possiblej.It is discussed below
How adjusting and optimizing statistical model parametrization vectorTo reach the purpose.
In order to obtain position of the three-dimensional statistical model summit on two dimensional surface, it is necessary to camera parameterAnd model parameterization
VectorThe initial value of the two parameters estimated and is manually entered by operator, and then two dimensional image is represented by
In order to weigh by vectorThe gap of statistical model real human face corresponding with actual picture after parametrization, definition
One departure function E=for weighing the gap | | Iinput-Imod||2, formula expression is with input picture feature point set and renders figure
The gap of the faceform of generation and real human face model are weighed as the Euclidean distance quadratic sum of feature point set, deviation is smaller,
Effect is better.In order to ask for departure function minimum using gradient descent method adjusting parameterization vectorHere it is false
If camera parameter is estimated correctly, soKeep constant.
λ in above optimization process is to approach the factor, and it reflects the speed of approaching of optimization process, and value is bigger, and speed is got over
Greatly, typically drawn by experience.But if the value is too small optimization process can be caused very long, it is also easy to be absorbed in local extremum.If
Value selection is larger, can not obtain more accurate model, it is also possible to cause departure function not restrained.In order to avoid case above
Occur, it is necessary to which dynamic adjustment approaches the factor.Used here as following strategy：
(1) initial time show that one larger is approached factor lambda by experienceinit；
(2) λ '=k λ are reduced if gradient direction changes during in iteration, wherein k is the ratio reduced every time
Example, typically takes 0.7~0.8.
Claims (4)
- A kind of 1. remains of the deceased three-dimensional rebuilding method based on direct picture, it is characterised in that：This method is in advance based on three-dimensional face number Three-dimensional face statistical model is established according to storehouse, and two-dimension human face spy is obtained using SDM Algorithm for Training based on two-dimension human face database Sign point parameter model；When shining into row human face rebuilding based on individual front, extracted first by two-dimension picture characteristic point parameter model Human face characteristic point；Then, according to three-dimensional face statistical model, energy function is entered using Studying factors self-adaption gradient descent method Row iteration optimizes, and obtains statistical model parametrization vector, carries out human face segmentation according to parametrization vector, that is, obtains two-dimension human face figure Three-dimensional face model corresponding to piece.
- 2. the remains of the deceased three-dimensional rebuilding method according to claim 1 based on direct picture, it is characterised in that：Use face Full face, the characteristic point for identifying and demarcating on human face photo, based on this group of Feature Points Matching three-dimensional face model space most Excellent solution, and ultimately generate three-dimensional face model.
- 3. the remains of the deceased three-dimensional rebuilding method according to claim 1 based on direct picture, it is characterised in that：Described use Two-dimension picture characteristic point parameter model extraction human face characteristic point is quick and precisely detected and identified on human face photo using SDM algorithms Characteristic point.
- 4. the remains of the deceased three-dimensional rebuilding method according to claim 1 based on direct picture, it is characterised in that：Described use Studying factors self-adaption gradient descent method is iterated optimization to energy function, obtains statistical model parametrization vector and refers to repeatedly Used gradient descent method is by improvement, the adaptive optimization side of Studying factors during generation optimization energy function Method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710793207.2A CN107564097B (en) | 2017-09-05 | 2017-09-05 | Three-dimensional reconstruction method for remains based on front images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710793207.2A CN107564097B (en) | 2017-09-05 | 2017-09-05 | Three-dimensional reconstruction method for remains based on front images |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107564097A true CN107564097A (en) | 2018-01-09 |
| CN107564097B CN107564097B (en) | 2021-02-05 |
Family
ID=60979174
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710793207.2A Active CN107564097B (en) | 2017-09-05 | 2017-09-05 | Three-dimensional reconstruction method for remains based on front images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107564097B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108346175A (en) * | 2018-02-06 | 2018-07-31 | 腾讯科技（深圳）有限公司 | A kind of facial image restorative procedure, device and storage medium |
| CN110021058A (en) * | 2019-03-21 | 2019-07-16 | 深圳点猫科技有限公司 | A kind of three-dimensional voxel modeling method and system convenient for juvenile's operation |
| CN110774583A (en) * | 2019-10-25 | 2020-02-11 | 上海轩林信息技术有限公司 | Modeling method for assisting in shaping of remains by color 3D printing and application of modeling method |
| CN111091624A (en) * | 2019-12-19 | 2020-05-01 | 南京大学 | Method for generating high-precision drivable human face three-dimensional model from single picture |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104794439A (en) * | 2015-04-10 | 2015-07-22 | 上海交通大学 | Real-time approximate frontal face image optimizing method and system based on several cameras |
| CN105844276A (en) * | 2015-01-15 | 2016-08-10 | 北京三星通信技术研究有限公司 | Face posture correction method and face posture correction device |
| CN106447763A (en) * | 2016-07-27 | 2017-02-22 | 扬州大学 | Face image three-dimensional reconstruction method for fusion of sparse deformation model and principal component regression algorithm |
-
2017
- 2017-09-05 CN CN201710793207.2A patent/CN107564097B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105844276A (en) * | 2015-01-15 | 2016-08-10 | 北京三星通信技术研究有限公司 | Face posture correction method and face posture correction device |
| CN104794439A (en) * | 2015-04-10 | 2015-07-22 | 上海交通大学 | Real-time approximate frontal face image optimizing method and system based on several cameras |
| CN106447763A (en) * | 2016-07-27 | 2017-02-22 | 扬州大学 | Face image three-dimensional reconstruction method for fusion of sparse deformation model and principal component regression algorithm |
Non-Patent Citations (3)
| Title |
|---|
| VOLKER BLANZ等: "《A Morphable Model For The Synthesis Of 3D Faces》", 《26TH INTERNATIONAL CONFERENCE ON COMPUTER GRAPHICS AND INTERACTIVE TECHNIQUES》 * |
| XUEHAN XIONG等: "《Supervised Descent Method and its Applications to Face Alignment》", 《2013 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION（CVPR）》 * |
| 邓秋平: "《基于单幅照片与统计模型的三维人脸重建》", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108346175A (en) * | 2018-02-06 | 2018-07-31 | 腾讯科技（深圳）有限公司 | A kind of facial image restorative procedure, device and storage medium |
| CN108346175B (en) * | 2018-02-06 | 2023-10-20 | 腾讯科技（深圳）有限公司 | Face image restoration method, device and storage medium |
| CN110021058A (en) * | 2019-03-21 | 2019-07-16 | 深圳点猫科技有限公司 | A kind of three-dimensional voxel modeling method and system convenient for juvenile's operation |
| CN110021058B (en) * | 2019-03-21 | 2023-09-26 | 深圳点猫科技有限公司 | Three-dimensional voxel modeling method and system convenient for child operation |
| CN110774583A (en) * | 2019-10-25 | 2020-02-11 | 上海轩林信息技术有限公司 | Modeling method for assisting in shaping of remains by color 3D printing and application of modeling method |
| CN111091624A (en) * | 2019-12-19 | 2020-05-01 | 南京大学 | Method for generating high-precision drivable human face three-dimensional model from single picture |
| CN111091624B (en) * | 2019-12-19 | 2021-09-28 | 南京大学 | Method for generating high-precision drivable human face three-dimensional model from single picture |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107564097B (en) | 2021-02-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN101404091B (en) | Three-dimensional human face reconstruction method and system based on two-step shape modeling | |
| CN109978930A (en) | A kind of stylized human face three-dimensional model automatic generation method based on single image | |
| CN104063702B (en) | Three-dimensional gait recognition based on shielding recovery and partial similarity matching | |
| CN101916454B (en) | Method for reconstructing high-resolution human face based on grid deformation and continuous optimization | |
| WO2015188684A1 (en) | Three-dimensional model reconstruction method and system | |
| WO2022001236A1 (en) | Three-dimensional model generation method and apparatus, and computer device and storage medium | |
| CN107564097A (en) | A kind of remains of the deceased three-dimensional rebuilding method based on direct picture | |
| CN108121950B (en) | Large-pose face alignment method and system based on 3D model | |
| CN106652015B (en) | Virtual character head portrait generation method and device | |
| KR20150113751A (en) | Method and apparatus for acquiring three-dimensional face model using portable camera | |
| CN112734890B (en) | Face replacement method and device based on three-dimensional reconstruction | |
| CN108154104A (en) | A kind of estimation method of human posture based on depth image super-pixel union feature | |
| KR20090092473A (en) | 3D Face Modeling Method based on 3D Morphable Shape Model | |
| CN110349247A (en) | A kind of indoor scene CAD 3D method for reconstructing based on semantic understanding | |
| CN107521090A (en) | It is a kind of that the incomplete method of remains face is precisely repaired based on 3D printing technique | |
| CN104954780A (en) | DIBR (depth image-based rendering) virtual image restoration method applicable to high-definition 2D/3D (two-dimensional/three-dimensional) conversion | |
| CN106652037B (en) | Face mapping processing method and device | |
| US20210375045A1 (en) | System and method for reconstructing a 3d human body under clothing | |
| CN106127818A (en) | A kind of material appearance based on single image obtains system and method | |
| CN109118455B (en) | Ancient human skull craniofacial interactive restoration method based on modern soft tissue distribution | |
| Zhang et al. | Data-driven flower petal modeling with botany priors | |
| CN111127642A (en) | Human face three-dimensional reconstruction method | |
| KR101116838B1 (en) | Generating Method for exaggerated 3D facial expressions with personal styles | |
| Ye et al. | 3d morphable face model for face animation | |
| Tena et al. | Fabricating 3D figurines with personalized faces |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |